You can now generate creatures directly within your project. Drop the open-source into your Agent, and start prompting what you want.
02 MID (budget 2 rounds | build: all parts, flat colour, ZERO new bones)
Output: every part — ears, horns, claws, eyes, fins, shell spikes, paws — riding bones LOW already placed, or anchored to volume surfaces. No material detail, no animation. MID's question is: do the key parts READ as what they are?
Author the WHOLE part set in one pass. Build once. Fix everything at once.
This is the single most expensive habit in the pipeline, so it gets the top of the card. Do not add parts one at a time and rebuild after each. Write every part, run one build, read the whole BLOCK list — the engine reports all failures at once, it does not stop at the first — fix them all in one edit, and build again.
The arithmetic on a 28-part body:
- the entire spec is a few percent of one build's output. The spec would have to be written dozens of times over to account for the run. Writing the spec is not the cost.
- MID takes ~130 turns for 40 elements — every part and volume touched 3.2 times on average. That is the cost.
A turn re-reads the whole conversation, which only grows. Touching 40 elements three times each is 120 round-trips over material that could have been written once and corrected once. Two passes is the target: author everything, then repair everything.
Mirrored structure is not authored twice, either. "mirror": ["LArm"] generates
the right side; hand-writing RArm alongside LArm doubles the work and the
chance of the two drifting apart. Use mirror for anything symmetric, and
joints_R when the pose (not the shape) differs.
MID SPENDS NO READER AT ALL
Both reads that used to live here are gone, and neither was replaced by a cheaper read — they were replaced by arithmetic that runs at build time.
The isolated part read is deleted. It failed every time it was tried — a part cut out of the body comes back as "a blob", "a stone", "a fist" — while the whole-body read of the same body had already named the parts. A part cut away from the body has lost the only context that told the reader what it was, so reading it there measures the crop, not the part.
The colour read is deleted too. It asked three things, and a machine answers all three off the palette in microseconds:
| the question | who answers it now |
|---|---|
| what colour is this creature | the palette. You wrote it. |
| which part draws the eye | value_order — every material sorted by OKLab lightness, brightest first, printed on every build. Whatever is at the top owns the eye. |
| which parts can you make out | contrast_adjacent — every part measured against the material of the thing it SITS ON. Under 0.10 OKLab they read as one mass. |
Eyeballing "value readability" costs hero renders and then a MID reader, to be told that a shin blade is the same colour as the leg under it. Both answers sit in the palette the whole time — a declared focal at L 0.975 tied for brightest with a claw at L 0.973 is exactly what those renders are being used to guess.
Both checks are ADVICE with numbers, not blocks: camouflage and deliberately subtle detail are real choices. What is not acceptable is paying a subagent to learn them.
part_exists still BLOCKS, and it costs nobody — it is a machine check on
the spec. If the brief named a chain and the spec has no such chain, the order
was not filled, and harness/brief.py says so at r1.
What a reader is still for, anywhere in this pipeline: what does this look like, which of these is punchier, and is this form legible. Those are judgments about resemblance and impact, and no amount of coordinate data performs them. Colour distance, value order, silhouette aspect, feature thickness and animation momentum are all arithmetic — none of them gets a reader again.
harness/partreads.py stays for a deliberate manual look at one part. No card
calls it.
Edit vocabulary (what a MID repair is allowed to be)
- swap the local shape: replace the primitive-ish mass with structured geometry — brow + muzzle + jaw instead of a head-sphere; knuckles + fingers instead of a fist-ball. No naked sphere / cube / cone may remain visible on a whitelisted part.
- fold and carve: concave sections,
sharpprofile breaks on the profile rows, a lowersmooth_angleon that volume (spec default 50°), plates and spikes that cut the outline.facetedon a VOLUME is a BLOCK (faceted_body): it shatters an 800-triangle torso into 800 shards, and AO bakes that mess into COLOR_0 where no relighting can reach it. Parts — fins, spikes, claws, crystal, armour — may still be faceted freely. - bend the pose: elbows/knees bend at MID — place the bend at its JOINT
(
jointspositions form the bind pose). Straight arms read as dead arms. Mirrored pairs may stagger viajoints_R(grow symmetric, pose asymmetric) — but only by a few centimetres. The twin's mesh is grown on the LEFT bone path and then translated onto the right joints, so a large offset shears the volume rather than posing it (a right thigh at 40% of the left's depth, width untouched — invisible from the left, invisible in the build log). Past ~10% of the limb's length it starts squashing andmirror_distortionwarns; past ~35% it BLOCKs. For a genuinely different pose, take that limb OUT ofmirrorand author it as its own chain.
Hands and feet — use the vocabulary, never improvise from spheres
Anything with hands gets "type":"hand" — palm, four fingers and an opposable
thumb, scaled from one size number; curl for relaxed/gripping, "fist": true for the folded fist with the thumb wrapped across. Field truth: every
attempt to improvise a hand from spheres and sticks shipped a mitten. Feet:
paw with "toes": 3..5 so they stop reading as bread loaves. Both take
mirrored: true, and the blind-read applies — a hand that reads as "a blob"
fails like any whitelisted part.
Joins — decide HOW every part meets the body, before placing it
A tusk that starts in mid-air, a beak hovering off the face, a trunk that reads as a bolted-on object: all one disease — the part was placed, but its RELATION to the body was never decided. For every hosted part, pick the join first:
- insert — the part SINKS into the mass (tusks into the jaw, horns into the
skull). Declare
"join":"insert"; the engine BLOCKs unless the base ring is ≥60% buried. Aim theoffsetback INTO the host volume, not at its surface. - extrude — the part GROWS out of the skin (a trunk from the face, a tail
spike from the tail). Declare
"join":"extrude"; the base centre must sit inside a body. Match the base radius to the local host radius so the skin flows into the part instead of stepping. - snap — the part LIES ON the surface (plates, scutes, brows). That is what
anchor+ conform already do; declare"join":"snap"and the engine insists the anchor exists. - place — deliberately detached (a floating rune, an orbiting shard). Rare;
say why in the spec
_notes.
Undeclared parts still get measured — a part_seat warn means a root may show.
And the join is also a COLOUR decision: a part that is FLESH of its host (trunk,
tail, brow) continues the host's material colour at its base; a hard colour
break at the junction reads as equipment, not anatomy. Check the junction in the
hero render, not just the part in isolation.
Do (layout)
- Signature presence: silhouette share is a BUDGET check — the signature part hits its declared share (judge measures by material name); nothing else needs a share test.
- 6:3:1 lands: geometry budget follows the hierarchy; the 6-part gets 6-level structure.
- Busy-vs-calm: on one part, one edge detailed, the other long and clean.
- Declared connections touch: masses that must read as one body declare
"touch": [["torso","tail"]]— the engine BLOCKS if they don't overlap. - Read the warns:
part_overlaplines flag parts sitting inside other parts (the self-intersecting-fist class). A warn is a measure, not a law — look at the render and decide.
Don't (blood of previous runs)
- Mirrored inward-tilted parts cross at the midline: length × sin(tilt) vs the left-right gap — do the arithmetic BEFORE placing horns/fins.
- Parts must not cover focals: a horn base once covered the eye. After placing anything big, re-check the focal view.
- Ride the right bone: jaw/tongue ride the skull, or they detach in motion.
- Every part must TOUCH its host —
part_attachmentBLOCKs a part whose nearest point still stands clear of the host surface. Tusks, trunks and plates used to float with the build reporting all green, and only a human eye caught them. The first third of a root is meant to be buried; that is what hides the seam. A part that cannot name what it touches is not attached. - Eyes/ears anchor to the HEAD volume surface:
"anchor": {"chain":"head","t":0.40,"around":62}—aroundin degrees from the top of the section (0=spine, 90=side, 180=belly). Wolf-class eyes sit at 55–70. - On an anchored plate the host surface wins by default: the plate is snapped
onto the surface normal and the direction you wrote survives as a reported
difference.
"conform": falsewhen that written direction was deliberate.
Stage-end gate
- Machine: build green + MID-stage claims (
part_exists/part_visible/part_signature/share_hierarchy/focal_contrast). - Whitelist blind-reads all passed.
- Human look #2: focal distribution right? part-to-body seams clean?
- Pass = part layout locks.